MiniMax H3 Keeps Generating the Same Face: How to Get Distinct Characters

AI Overview

Why does MiniMax H3 keep generating the same face?

Broad character prompts can repeatedly land on a familiar-looking face even when the seed changes. The model still has to invent identity, so vague wording leaves too much of that decision unconstrained.

Will a random seed always create a different person?

No. A new seed changes the sample, but it does not guarantee a new facial structure. If the prompt describes the same generic role, several outputs can still look like close relatives.

Should I use a reference image to create more face diversity?

Use references after you choose the identity you want to preserve. For initial casting diversity, text-only batches with controlled identity variables make it easier to compare genuinely different candidates.

How do I know whether two generated faces are too similar?

Compare stable traits rather than hair or wardrobe: face shape, brow, eye spacing, nose bridge, jaw, age cues and skin texture. Review neutral front views before judging moving clips.

What “Same Face” Actually Means

When MiniMax H3 keeps generating the same face, the failure is usually not a literal duplicate pixel for pixel. It is a repeated casting pattern: similar age, symmetrical oval face, narrow jaw, familiar eye spacing, smooth skin and the same editorial expression. Hair, clothing and location may change enough to disguise the repetition at first. Once the clips are placed side by side, the characters feel related or interchangeable.

That is a different search intent from keeping one chosen character consistent. Consistency asks H3 to protect identity across shots. Face diversity asks text-to-video to invent clearly different identities at the casting stage. Do not solve the second problem by adding a random reference image; that only locks another identity before you know whether it fits.

Editorial concept image: different hair and clothing can hide the fact that brow, eye spacing, nose and jaw still repeat.

Separate identity repetition from ordinary continuity

First decide what is repeating. If the same face appears in several unrelated text-only generations, you have a diversity problem. If one face changes inside a single clip, you have identity drift. If two named people merge during an interaction, you have subject binding or occlusion failure. Each needs a different fix.

The official H3 system exposes text-only FL2VA generation and multimodal Ref2VA generation as different task families. That distinction matters: text-only generation creates the person from language, while reference generation can inherit a selected visual identity. The model documentation does not promise that different seeds will produce a balanced casting pool, so treat diversity as something to test and direct.

Check whether the prompt is secretly asking for one archetype

Prompts such as “beautiful young woman,” “handsome cinematic man,” or “stylish creator” describe a role and aesthetic, not an individual. Repeating the same role, lens, beauty light and mood can keep steering generations toward the same visual center. Extra quality words do not add identity information.

Also check negative prompts. Removing wrinkles, asymmetry, textured skin, unusual hairlines or strong age cues can erase the traits that make candidates different.

Build a Clean Diversity Test

Do not diagnose this from two attractive thumbnails. Build a small casting test that separates facial identity from scene styling. Generate six short neutral clips using the same duration, frame rate, aspect ratio, camera position, wardrobe category, background and lighting. Change only the identity description and seed policy. A static medium close-up with a small head turn is enough; dramatic action makes comparison harder.

Editorial concept image: equal lighting, framing and wardrobe make facial differences easier to judge.

Use a six-candidate casting matrix

Give every candidate one deliberately different profile. Vary several stable traits together instead of changing only hair color.

Use respectful, observable descriptors and avoid guessing protected traits. This production matrix simply shows whether prompt changes reach the rendered face.

Save a neutral comparison frame

Export one front or three-quarter frame from the same moment in each clip. Crop each to equal head size and review without filenames. Score face shape, brow, eyes, nose, jaw and age cues from 0 to 2: zero means clearly different, one means partly similar, and two means strongly similar. Two candidates that share four or more traits should not both pass a campaign casting set unless the resemblance is intentional.

Keep the full clips too. A still can look distinct while motion pulls the face toward a familiar expression. Review the first, middle and final seconds before approval.

Rewrite the Prompt to Define a Different Person

The fastest fix is not a longer prompt. Replace broad attractiveness language with a compact identity block and let the rest of the prompt describe action. The MiniMax H3 generator gives you a hosted place to run the same controlled test without rebuilding a local graph.

Use one identity block and one shot block

Copy this structure and change only the identity line between candidates:

Identity: a fictional middle-aged ceramic artist with a long angular face, strong brow, prominent nose bridge, short salt-and-pepper curls, natural skin texture and a calm unsmiling expression. This person must not resemble prior candidates.

Shot: medium close-up in a neutral daylight studio. The subject looks toward camera, turns slightly left, then holds still. Static 50 mm-style framing, soft side light, plain charcoal crewneck, no beauty retouching, no text or logo. Quiet room tone only.

“Must not resemble prior candidates” is useful as an intention but cannot compare against images the model never received. The concrete traits do the real work. Keep the action, lens language and light identical so you can tell whether the identity line changed the person.

Editorial concept image: change the performer definition while holding the production setup constant.

Remove prompt conflicts

A prompt can ask for an older face and then undo it with “youthful,” “flawless,” “smooth skin” and “beauty commercial.” It can request a broad jaw, then introduce a reference whose subject has a narrow face. Resolve those conflicts before spending more seeds.

One main action and one camera move are enough for casting. Once a candidate is approved, restore the real scene, costume, performance and camera plan. The Ref2VA versus FL2VA guide explains when text or keyframes should invent the shot and when references should take control.

Use Seeds and References Deliberately

A seed is a reproducibility control, not a casting instruction. Change it to explore a prompt, but do not assume a new integer means a new identity. Record the prompt, task family, checkpoint, seed, duration and workflow version so a useful candidate can be reproduced.

Batch first, then lock the winner

Generate the controlled six-candidate batch without identity references. Select one candidate on facial fit, motion stability and suitability for the role—not simply the prettiest first frame. Then save a neutral portrait and a useful full-body or three-quarter image as the approved identity pack.

Move to Ref2VA only after approval. Give each visual reference a clear job: one for face, one for body or wardrobe, and perhaps one for environment. The character replacement workflow shows how to keep a chosen performer while preserving an existing shot's motion and staging.

Play a documented MiniMax H3 reference-to-video output

Documented MiniMax H3 Ref2VA output: use reference mode after casting to inspect whether the selected identity survives motion, camera changes and audio together.

Do not mix casting and continuity in one batch

If every candidate uses the same portrait reference, repeated faces are expected. A motion reference with a visible performer may contribute more than motion unless its role is clear. Test character invention separately, then combine approved inputs.

For scenes with two people, label them and keep their physical roles distinct. The two-person MiniMax H3 workflow covers subject IDs, eyelines and turn-taking so one identity does not bleed into the other.

Troubleshoot Persistent Lookalikes

If detailed identity blocks still produce lookalikes, find the first stage where they appear. Compare the 768p base result before regeneration, upscaling, face enhancement or interpolation. Restoration can normalize skin and facial detail so aggressively that different inputs converge.

Change one variable in this order

1. Remove beauty and quality adjectives that imply one editorial archetype.
2. Strengthen structural traits: face shape, jaw, nose bridge, brow and age cues.
3. Simplify action and camera movement.
4. Change seed while keeping the revised identity block fixed.
5. Compare base output before any enhancement.
6. If the role is still wrong, cast with a new approved reference rather than burning more random seeds.

Do not jump straight to a face-restoration pass. Restoration can improve detail, but it cannot prove that the underlying identity is different. Judge diversity before polish.

Know when the apparent duplicate is a review problem

Small faces in wide shots hide distinguishing features. Backlight, sunglasses, hair over the face and fast motion can make different people look similar. If neutral close-ups are distinct, solve the wide shot with framing and light rather than recasting.

Review a Diverse Casting Set

Review the candidates as a set, not one clip at a time. A strong individual result can still duplicate another approved character. Put neutral frames on one board, hide seed numbers and prompt names, and ask a reviewer to group apparent relatives. Any accidental cluster returns to the casting test.

Editorial concept image: approve the whole cast together so repeated facial patterns do not slip through separate reviews.

Keep a reusable casting record

For each approved character, save the neutral face frame, full-body reference, exact identity block, seed, task mode, accepted clip and a short “do not converge toward” note. Do not use a real person's likeness without permission. Keep fictional or authorized references clearly separated from discarded generations.

This is where Seedance Agent reduces coordination work: attach the casting board to the shot plan, keep each reference role explicit, compare candidates, approve one identity and rerun only the failed character or shot. The point is not to automate taste; it is to preserve the evidence behind the casting decision.

Conclusion

When MiniMax H3 keeps generating the same face, treat it as a casting-diversity problem rather than an identity-consistency problem: remove generic beauty language, define several structural traits, run a controlled six-candidate test, review neutral frames, and use seeds as exploration controls rather than guarantees. Approve a genuinely distinct performer before switching to references, then keep casting records and continuity checks attached to the production. To plan the cast, compare takes and carry approved identities into later shots without losing the decision trail, build the project with Seedance Agent.